3D U-Net vs. 2D U-Net for Brain MRI Segmentation: A Comparative Study on Accuracy and Efficiency

DOI: https://doie.org/10.10399/JBSE.2025237518

Astha Singh , Aditi Sharma

Keywords:

Brain MRI, segmentation, deep learning, 3D U-Net, 2D U-Net, medical imaging, gray matter, white matter, cerebrospinal fluid.

Abstract:

Precise segmentation of brain tissues in MRI scans is vital for diagnosing neurological disorders and planning treatments. While 2D U-Net models are computationally efficient, they process images slice by slice and can miss spatial context between slices, which is crucial for accurate segmentation of complex brain structures. This study compares 2D and 3D U-Net architectures using the iSeg-2017 infant brain MRI dataset to analyze their performance in segmenting gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF). Evaluation metrics including Dice coefficient, sensitivity, specificity, and precision indicate that the 3D U-Net consistently achieves higher scores, with Dice values reaching up to 0.950 for CSF. Preprocessing techniques like skull stripping, intensity normalization, and patch-based training contributed to improved model accuracy. Despite the improved results from the 3D model, the performance gains over 2D remain modest, suggesting that 2D models could still be suitable in situations with limited computational resources. Future work will focus on expanding the dataset size, validating the models on external scans, and exploring hybrid architectures to enhance segmentation precision and clinical applicability.

PDF